614 8 AUGUST 2014 • VOL 345 ISSUE 6197 sciencemag.org SCIENCE

The brain chip

PH

OT

O:

DE

AN

NE

FIT

ZM

AU

RIC

E

On a computer monitor in IBM’s

brain lab here, a video taken from

atop a tower on the Stanford Uni-

versity campus shows a steady

stream of cars, bikes, buses, trucks,

and pedestrians as they come and

go. As each shape enters the scene,

it’s briefly surrounded by a splash

of color: purple for cy-

clists, green for pedestrians, dark

blue for cars, sky blue for trucks,

and yellow for buses. The colors

signal the judgments made by a

postage stamp–sized computer

chip, which surveys the ever-

changing scene and identifies each on-screen

target. “It gets almost everything right,” says

Dharmendra Modha, an electrical and com-

puter engineer who leads the project at IBM’s

Almaden Research Center in the hills beyond

Silicon Valley. A bicyclist entering from the

right is quickly wrapped in purple. But as the

cyclist stops, dismounts, and starts to walk

his bike, the color shifts to pedestrian green.

Modha smiles. “We got a little lucky with that

one,” he says.

Easy for a human, such pattern recog-

nition is a tour de force for a computer.

On page 668 of this issue, Modha and col-

leagues at five IBM research cen-

ters and Cornell University describe

the chip responsible for it: the

first-ever production-scale “neuro-

morphic” computer chip designed

to work more like a mammalian

brain than like the processors in a

laptop or smart phone. The new chip, called

TrueNorth, marks a radical departure in chip

design and promises to make computers

better able to handle complex tasks such as

image and voice recognition—jobs at which

conventional chips struggle.

Progress in brain-inspired computing

has been building for several years. Several

U.S. and European labs are working on dif-

ferent versions of the technology. But out-

siders say TrueNorth has the potential to

propel neuromorphic computing from an

alluring research endeavor to a real-world

technology.

TrueNorth contains 5.4 billion transistors

wired together to form an array of 1 million

digital “neurons” that talk to one another

via 256 million “synapses.” Like the brains

of organisms, this neural network archi-

tecture accomplishes complex tasks such

as pattern recognition far more efficiently

than conventional chips can. “It’s a tremen-

dous achievement,” says Wei Lu, an electri-

cal engineer and computer scientist at the

University of Michigan, Ann Arbor. Horst

Simon, a computer scientist and deputy

director of the Lawrence Berkeley National

Microprocessors modeled on networks of nerve cells promise blazing speed at incredibly low power—if they live up to hopes

By Robert F. Service, in San Jose, California

To hear a podcast with author Robert F. Service, see http://scim.ag/pod_6197.

PODCAST

Published by AAAS

on

Dec

embe

r 3,

201

4w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

o

n D

ecem

ber

3, 2

014

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Dec

embe

r 3,

201

4w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

8 AUGUST 2014 • VOL 345 ISSUE 6197 615SCIENCE sciencemag.org

NEWS | FEATURES

Laboratory in California, agrees. “This is

a qualitative breakthrough to go from one

computing paradigm to the next,” he says.

TO US, PERCEPTION seems so simple. We

peer out the window of an office building

and easily distinguish between a person on

a bike and one on a skateboard. When we

read a sentence that mentions a calico and

a tabby, we understand that both are cats

without having to be told. And with little ef-

fort we pick out the voice of the person we

are talking to at a noisy cocktail party. By

comparison, modern computers, for all their

powers of calculation, remain infants at such

tasks. State-of-the-art algorithms can crack

these challenges, but they require massive

computing power. Google, for example, re-

cently demonstrated a setup capable of rec-

ognizing cats and human faces in video clips.

But the task required 16,000 processing

chips and about 100 kilowatts of power. Our

brains, by contrast, use just tens of watts.

Today’s chips are based on the same ar-

chitecture that was developed 7 decades

ago by the Hungarian-born polymath John

von Neumann. In 1945, von Neumann laid

out modern computing’s basic design, with

separate processing, memory, and control

units. The architecture excels at perform-

ing sequences of logical operations and is

ideal for crunching numbers and running

spreadsheets and word processors. But it

struggles when trying to integrate and pro-

cess large amounts of data, as vision and

language processing demand.

The difficulty arises in the way conven-

tional chips carry out a task. To do anything

useful, they must pull data in from stored

memory, manipulate it, and send the result

back to storage before tackling the next oper-

ation. Moving all that data back and forth de-

mands power and creates traffic bottlenecks.

For decades, engineers have compensated

by shrinking the transistors, communica-

tion lines, and other devices on chips. That

shortened the distance data needed to travel,

reduced the power demands of individual

devices, and sped them up.

But this strategy is all but tapped out. In-

dividual device features in the latest chips

are as small as 14 nanometers across—the

width of fewer than 100 atoms, and close to

the limits set by physics. To keep computer

power on its upward trajectory, makers have

resorted to tiling multiple processor chips

side by side. That approach compounds the

need to shuttle data—and the challenge of

matching the skills of biology.

In 2012, for example, Modha and his col-

leagues used an IBM supercomputer called

Sequoia at Lawrence Livermore National

Laboratory in California to simulate the

network communication in a human-scale

brain. The simulation used conventional

circuitry programmed to emulate the com-

munication among 500 billion neurons

and 100 trillion synapses. All of Sequoia’s

1.5 million processor chips and 1.5 petabytes

(1.5 quadrillion bytes) of memory were de-

voted to the task—and even so, the simula-

tion ran at only 1/1500 the speed of a real

brain. If the simulation had been scaled

up to keep pace with actual “wetware,” it

would have required 12 gigawatts of power,

Modha says—the power consumption of Los

Angeles and New York City combined.

Biology takes a different approach. Indi-

vidual neurons in our brains communicate

with thousands of other neurons through

chemical signals at connection points called

synapses. When the combined chemical sig-

nals from a neuron’s partners exceed a cer-

tain threshold, it fires, creating an electrical

spike that triggers it to pass chemical signals

to other partners. And the communication

ripples through the brain’s network, which

in humans includes an estimated 100 billion

neurons and 100 trillion synapses.

The setup needs no program: Connec-

tions to other neurons that fire regularly

are reinforced, whereas those that don’t are

pared away. In that way, the architecture

itself learns. The strategy is also efficient,

because once a nerve fires, it “forgets” all

about the inputs that pushed it over the

threshold and passes on only the fact that

it fired. And it enables the brain to distrib-

ute information processing tasks. Separate

groups of neurons in the visual cortex, for

example, respond to horizontal and verti-

cal edges in a scene and pass those per-

ceptions to other neurons that integrate

the information. This midlevel processing

minimizes the need to shuttle data from

one place to another.

The upshot, says neuromorphic chip ex-

pert Gill Pratt, is that whereas conventional

computer processors base their performance

mainly on speed, the wetware in a brain re-

lies on the complexity of its network—which

lowers energy consumption. That’s a sur-

vival advantage, says Pratt, who heads the

SyNAPSE (Systems of Neuromorphic Adap-

tive Plastic Scalable Electronics) program

at the Defense Advanced Research Projects

Agency in Arlington, Virginia, which funds

the IBM effort and a companion program at

HRL Laboratories LLC in Malibu, Califor-

nia. “In nature, the most precious resource

for most animals is food,” he explains.

Nature’s emphasis on parallel process-

ing and complexity is at the heart of most

attempts at neuromorphic computing. The

term was coined in the 1980s by California

Institute of Technology electrical engineer

Carver Mead to describe efforts to mimic

neurological architectures using analog

computer circuits—devices that, like neu-

rons, send signals only after inputs from

their neighbors reach a predetermined

threshold. But today, researchers apply the

term more broadly to a range of analog, dig-

ital, hybrid, and even software systems. Like

wetware, all neuromorphic systems distrib-

ute processing and memory tasks broadly

across the chip to minimize the need to

ferry data back and forth to central logic

and memory hubs.

One effort, run by Narayan Srinivasa at

HRL Laboratories, follows some of Mead’s

One neuromorphic chip can spot patterns, such as

clues that distinguish cyclists from pedestrians, as

handily as an array of power-hungry processors.

Dharmendra Modha’s team at IBM aims to give com-

puters the perception skills of biological organisms.

CR

ED

IT:

DE

AN

NE

FIT

ZM

AU

RIC

E

Published by AAAS

NEWS | FEATURES

616 8 AUGUST 2014 • VOL 345 ISSUE 6197 sciencemag.org SCIENCE

PH

OT

OS

: D

EA

NN

E F

ITZ

MA

UR

ICE

inspiration with analog circuitry. Srinivasa’s

team has developed a silicon chip with ana-

log circuitry containing 576 neurons and

73,000 synapses. In June, Srinivasa’s team

showed that the chip could interpret visual

signals well enough to pilot a palm-sized

helicopter through a building. The chip

picked up signals from an imaging sensor and

used that information to determine whether

it was in a new room or one that it had flown

through previously. Because the setup uses

analog circuitry, Srinivasa says, the chip can

strengthen connections over time and thus

learn and improve its performance.

Another project is being run by Stephen

Furber of the University of Manchester in

the United Kingdom. Known as SpiNNaker

( S p i k i ng N e u r a l N e t w o r k A r c h i t e c t u r e ) ,

Furber’s system is a supercomputer con-

structed from conventional digital-based

low-power computer chips, the sort com-

monly found in smart phones. SpiNNaker

now consists of 20,000 chips, each of which

represents 1000 neurons. This fall, Furber

says, he expects that number will rise to

100,000 chips representing 100 million neu-

rons, and eventually a 1-million-chip system

representing 1 billion neurons—about 1% of

the neurons in the human brain. Although

similar in concept to the Sequoia simulation,

SpiNNaker’s design is expected to model

brain activity at speeds matching biology.

Furber says he and his colleagues aren’t

building SpiNNaker for particular applica-

tions. For now, he says, the machine will

serve as a testbed for computer scientists

and neuroscientists to model brain func-

tion in a way that connects

their small-scale knowl-

edge of neurons with the

global insights of brain

imaging and psychology.

With SpiNNaker, “you can

ask very detailed questions

that are very difficult to

ask in biology,” Furber says.

IBM’S TRUENORTH ALSO

uses conventional digital

devices. But in this case,

Modha and his colleagues

wired them in an entirely

new hardware architec-

ture. So far, that’s enabled

them to represent 16 times

as many neurons on a

single chip as previous ef-

forts could, thus speeding

up its parallel processing at

low power.

Modha and his col-

leagues unveiled their first

attempt in 2011: a pencil

tip–sized chip containing

256 neurons and 262,000 synapses, a unit

they refer to as a core. Their current chip

contains an array of 4096 cores. “This is the

largest chip IBM has ever made by a factor

of 3,” Modha says. And because they used

more advanced chipmaking techniques to

build it, each of TrueNorth’s cores is just

1/15 the size of the previous generation and

consumes 1/100 as much power. That makes

it more than 1000 times as efficient as chips

made with the conventional architecture.

Whereas a typical chip sucks down 50 to

100 watts per square centimeter of chip

space, TrueNorth sips just 20 milliwatts for

the same area of circuitry.

Modha and his colleagues are already

working to build on their success. They are

testing arrays of up to 16 TrueNorth chips

and are getting ready to push beyond that.

Because each chip already represents an ar-

ray of cores wired together, it’s a straight-

forward task to tile more and more chips

together to increase the computational

power. On the drawing board are collections

of 64, 256, 1024, and 4096

chips. “It’s only limited by

money, not imagination,”

Modha says.

That’s just the hardware.

To control it, Modha and

his colleagues have also

created a novel program-

ming language—a type of

machine language for syn-

aptic chips. Modha says

engineers have already

used it to write hundreds

of simple software rou-

tines called corelets that

t e l l T r u e N o r t h ’s n e u r a l

network how to carry out

c o m m o n c o m p u t a t i o n a l

tasks. Vision corelets that

spot features such as ver-

tical and horizontal edges

helped TrueNorth distin-

guish the cars, cyclists, and

pedestrians in the brain

lab video, Modha says.

Equipped with an array of

these basic functions, pro-

grammers should be able to rework conven-

tional, inefficient neural network programs

to run on the IBM chips. Thousands of such

programs already exist for carrying out tasks

such as visual and auditory recognition. The

IBM chip should be able to run them far

more quickly while drawing less power.

IBM hopes to commercialize the chips

“sooner rather than later” and is looking

into partnerships with other companies,

Modha says. Meanwhile, he says, IBM

plans to give computer scientists access to

the chips to explore the cornucopia of new

applications they should make possible.

Berkeley Lab’s Simon thinks that in time

the chips could lead to a new generation of

power-efficient supercomputers. “There’s a

huge potential here for addressing differ-

ent types of calculations that are currently

done on [conventional chips] but less effi-

ciently,” Simon says.

Modha’s team has dreamed up some

applications of its own. One of them, dis-

played in the team’s brain lab across the

room from the video monitor tracking

Stanford traffic, imagines an advanced ver-

sion of Google Glass that processes visual

information and communicates it to visu-

ally impaired wearers. A blind person, for

example, might receive auditory cues to

avoid objects in her path. In a second exam-

ple, called Tumbleweed, a robot outfitted

with multiple sensors rolls its way around a

dangerous environment, such as the inside

of a damaged nuclear reactor, and beams

back visual, temperature, chemical, and

radiation data. In the nearer term, Furber

suggests, neuromorphic chips could be in-

tegrated into smart phones to improve their

visual and voice recognition.

To encourage such applications, IBM

has set up a virtual school, called Synapse

University, where computer scientists and

researchers can learn how to program the

new chips to do whatever they want. “If

IBM can do that, I’m sure lots of people

will have fun with it and do real science,”

Furber says. And eventually, perhaps, com-

puters will begin to emerge from their in-

fancy in carrying out everyday tasks that

we take for granted. ■

Neuromorphic chips could lead to

rolling robots (top) that beam back

data from hazardous environments

or to navigation aids for the blind.

Each TrueNorth chip contains

5.4 billion transistors wired into

an array of 1 million neurons and

256 million synapses. Efforts are

already under way to tile more

than a dozen such chips together.

Published by AAAS